Support mtl-2.3-rc4

[andreasabel]

Dependencies breaking with mtl-2.3:

• equivalence Breakage with mtl-2.3-rc4 pa-ba/equivalence#14
• exceptions transformers-0.6 support ekmett/exceptions#82
• constraints Build failure with mtl-2.3 ekmett/constraints#110 (Fixed by constraints-0.13.3)

Problems in our code:

• transformers-0.6, enabled by mtl-2.3 breaks compilation of instance MonadTrans TCMT.

Context:

Release of mtl-2.3 is imminent, see

• https://github.com/haskell/mtl/releases/tag/v2.3-rc4
• https://mail.haskell.org/pipermail/libraries/2021-November/031523.html (quoted below)

Announcement mail by Emily Pillmore

Chessai and I are excited to announce a release candidate for mtl: mtl-2.3-rc3!

Timeline

The timeline for release will be as follows:

• This announcement marks the start of the timeline

• We will give 2 weeks of testing before considering release

• If no major issues are filed before then, mtl-2.3-rc3 will be released as mtl-2.3, and if issues are found, they will be amended, and a new release candidate will be announced, resetting the 2 week period.

To test mtl-2.3-rc3 for yourself, please feel free to add the following to your cabal.project:

source-repository-package
  type: git
  location: https://github.com/haskell/mtl.git
  tag: 5d0f62b8007bb96e49f36a5544741cfe96a97130

or, if you're a stack user add this entry to your extra-deps:

- git: https://github.com/haskell/mtl.git
  commit: 5d0f62b8007bb96e49f36a5544741cfe96a97130

And make sure to adjust all bounds/allow-newer accordingly. Please note that this release of mtl is a full major version release, and will be the last of the 2.x series before work begins on updating the mtl class hierarchy.

Changelog

• Add instances for Control.Monad.Trans.Writer.CPS and Control.Monad.Trans.RWS.CPS from transformers 0.5.6 and add Control.Monad.Writer.CPS and Control.Monad.RWS.CPS.

• Control.Monad.Cont now re-exports evalCont and evalContT

• Add tryError, withError, handleError, and mapError to

Control.Monad.Error.Class, and re-export from Control.Monad.Except.

• Remove Control.Monad.List and Control.Monad.Error

• Remove instances of deprecated ListT and ErrorT

• Remove re-exports of Error

• Add instances for Control.Monad.Trans.Accum and Control.Monad.Trans.Select ( http://control.monad.trans.select/ )

• Remove re-exports of Control.Monad, Control.Monad.Fix and Data.Monoid modules

I'd like to thank the many contributors who offered patches, tickets, and other help in the preparation of this release. We appreciate all of your help!

Happy hacking!

Emily

[andreasabel]

If mtl-2.3 is released as planned it will break all of hackage.

• mtl-2.3-rc3 breakages due to re-exports breaks haskell/mtl#102

Because of this, there could be delays. Esp. stackage nightly might not switch so quickly. Thus, I removed this from the 2.6.2.1 milestone.

[andreasabel]

The new RC4 breaks very little, but the removal of ErrorT will still affect us.
(EDIT: Rather, our dependencies.)

[andreasabel]

Seems we are hit by the following innovation in transformers-0.6:

0.6.0.0 Ross Paterson R.Paterson@city.ac.uk Jul 2021

• Added quantified constraint to MonadTrans (for GHC >= 8.6)

class (forall m. Monad m => Monad (t m)) => MonadTrans t where

Problem here:

agda/src/full/Agda/TypeChecking/Monad/Base.hs

Lines 4263 to 4276 in c7499fb

	instance MonadTrans TCMT where
	lift m = TCM $ \_ _ -> m
	-- We want a special monad implementation of fail.
	instance MonadIO m => Monad (TCMT m) where
	return = pure
	(>>=) = bindTCMT
	(>>) = (*>)
	#if __GLASGOW_HASKELL__ < 808
	fail = Fail.fail
	#endif
	instance MonadIO m => Fail.MonadFail (TCMT m) where
	fail = internalError

src/full/Agda/TypeChecking/Monad/Base.hs:4263:10: error:
    • Could not deduce (MonadIO m)
        arising from the superclasses of an instance declaration
      from the context: Monad m
        bound by a quantified context
        at src/full/Agda/TypeChecking/Monad/Base.hs:1:1
      Possible fix:
        add (MonadIO m) to the context of a quantified context
    • In the instance declaration for ‘MonadTrans TCMT’

Reported this problem upstream:

• https://hub.darcs.net/ross/transformers/issue/85
• https://gitlab.haskell.org/ghc/ghc/-/issues/19922#note_405825

[andreasabel]

According to my experience so far, Agda 2.6.2.1 could be relaxed to mtl < 2.4 if restricted to transformers < 0.6 (the latter is implicit in mtl < 2.3). Let's do this revision once mtl-2.3 is out.

[RossPaterson]

Would this be fixed if agda changes the instance to

#if __GLASGOW_HASKELL__ < 808
instance MonadIO m => Monad (TCMT m) where
#else
instance Monad m => Monad (TCMT m) where
#endif
    return = pure
    (>>=)  = bindTCMT
    (>>)   = (*>)
#if __GLASGOW_HASKELL__ < 808
    fail   = Fail.fail
#endif

and transformers limits the constraint to GHC >= 8.8?

[andreasabel]

@RossPaterson : Thanks for engaging! I will investigate your suggestion and get back to you.

[andreasabel]

@RossPaterson : Indeed, your fix simply works for all our GHC versions!, except for GHC 8.6.
(I was afraid the choice of superclass constraint of Monad (TCMT m) would affect more locations, but its really just that bit now (after my previous AMP-refactoring).)

So your suggested revision of transformers-0.6.0 would work for us.

[RossPaterson]

Glad to hear it. The change is in transformers-0.6.0.3.

[andreasabel]

Revision 3 of 2.6.2.1 to allow mtl-2.3 but restrict to transformers-0.5.*: https://hackage.haskell.org/package/Agda-2.6.2.1/revisions/

[phadej]

I'm very late to the party, but GHC-8.6 already had MonadFailDesugaring, so defining fail in Monad instance wouldn't been strictly required. Having default-extensions: MonadFailDesugaring in Agda wouldn't needed to make breaking transformers release.

[andreasabel]

@phadej: Thanks for pointing this out. Unfortunately, we didn't have the expertise to find the best solution back then. Indeed, I can simply relax the fix from GHC-8.8 to GHC-8.6 as the latter has MonadFailDesugaring on by default. I do not even need to add it to the Agda.cabal file (which would be a bit troublesome as it would have to be a default-extensions entry in a conditional).
I verified your suggestion in this PR:

• Re #5659: no need to exclude transformers-0.6.0.0/1/2 on GHC 8.6 #6357